Enhancing streamflow prediction in a mountainous watershed using a convolutional neural network with gridded data.

In this research, we demonstrate the effectiveness of a convolutional neural network (CNN) model, integrated with the ERA5-Land dataset, for accurately simulating daily streamflow in a mountainous watershed. Our methodology harnesses image-based inputs, incorporating spatial distribution maps of key environmental variables, including temperature, snowmelt, snow cover, snow depth, volumetric soil water content, total evaporation, total precipitation, and leaf area index. The proposed CNN architecture, while drawing inspiration from classical designs, is specifically tailored for the task of streamflow prediction. The model's performance, assessed during both the training and testing phases, demonstrates high accuracy, reflected quantitatively in metrics such as RMSE, MAPE, R², and NSE. Notably, the model exhibits enhanced accuracy in predicting lower flow rates, particularly in autumn and winter, as evidenced by an average RMSE of 2.02 m³/s for flows below 13.8 m³/s. In contrast, the model's precision decreases in high flow rate scenarios, predominantly in spring and early summer. The implementation of forward feature selection (FFS) has further optimized the model, pinpointing total evaporation and volumetric soil water as key parameters, thus enabling a more efficient model with accuracy comparable to the initial, more complex version. This research underscores the practical utility of an image-based approach using CNN models for streamflow prediction. Moreover, the adoption of the freely available and universally accessible ERA5-Land dataset highlights its effectiveness as a valuable and cost-efficient tool for streamflow prediction.